1

I am currently writing my own voice assistant in python using nltk for preprocessing and pytorch for processing the data. After lots of hours searching for any method, I can't find a way to extract the title of a song from other spoken text. So what I want to achieve is at example filtering "Numb" out of "Play numb by Linkin Park". Is this somehow possible with NLP or just using neural network and how?

PrimoDev23
  • 11
  • 1
  • Could you provide a minimal example of what you've already tried? – PirateNinjas Nov 22 '21 at 15:00
  • @PirateNinjas basically I just tried a simple feed forward neural net. My input layer had 20 inputs which corresponds to the text with some padding. As output I have used the indices of the song title (so start index and end index). But after training I juts get values < 0, so that doesn't seem to work. Also my method to convert strings to numbers is kinda weird. Just getting the characters values and putting them together. – PrimoDev23 Nov 22 '21 at 16:02

1 Answers1

-1

This is potentially quite a difficult problem to solve generally. As a first pass, you could try imposing some additional assumptions:

  1. The text passed to your "song name extractor" is perfectly translated from speech
  2. The user will follow a set format for requesting the song

If you make these assumptions the problem can be solved using regex, like so:

import re

# your input text
song_request = "Play numb by Linkin Park"

# search the input text for a matching substring
song_search = re.search("(?<=Play ).*(?= by)", song_request)

# if you get a match, extract the song title
if song_search:
    song_title = song_search.group()
else:
    song_title = ""  # just in case your assumption doesn't hold
PirateNinjas
  • 1,908
  • 1
  • 16
  • 21
  • Well with a fixed pattern that's possible yeah but since the assistant is in German there is more than one possible sentence pattern. Generally I would prefer not to use static patterns. I have no problem with building a neural net but my current tests don't look promising – PrimoDev23 Nov 22 '21 at 14:59